Machine Learning-Based Optimization for Enhanced Pullulan Recovery\nfrom Aureobasidium pullulans

The\nsolvent extraction protocol significantly influences the yield\nand purity of the exopolysaccharide recovery from cell-free broth.\nIn this study, methanol, ethanol, isopropanol, acetone, and PEG-6000\nwere compared for the amount of pullulan recovery (PR), sucrose equivalent\n(SE), and protein impurity (PI) of the precipitate. The PR (7.0 g\nL<sup>–1</sup>) and SE (0.45 g g<sup>–1</sup>) from\nacetone had been significantly better than others. Quadratic and GA-optimized\nartificial neural network (ANN) models, developed from a central composite\ndesign, accurately predicted PR (<i>R</i><sup>2</sup>: 0.996–0.998),\nSE (<i>R</i><sup>2</sup>: 0.961–0.985), and PI (<i>R</i><sup>2</sup>: 0.952–0.984) based on the pH, incubation\ntime, and solvent-to-broth volume (S/B) ratio. Individually optimized\nPR (10.53 g L<sup>–1</sup>), SE (0.64 g g<sup>–1</sup>), and PI (1.05 mg g<sup>–1</sup>) from quadratic models are\ncomparable to those obtained from ANN models. Multiobjective optimization\nwith equal weighting suggests a moderate PR (quadratic model: 8.14\ng L<sup>–1</sup>, ANN: 6.77 g L<sup>–1</sup>) while\nmaintaining SE and PI close to their optimal values. However, the\nrelative importance of these objectives should be ascertained based\non the intended application.